Reservoir computing with brain organoids uses a living organoid as the “reservoir”: a fixed, complex network that transforms input signals into rich patterns of activity, while only a simple digital readout is trained. In Brainoware (Cai et al., Nature Electronics, 2023), an organoid on an electrode array helped identify speakers from Japanese vowel recordings and predict a chaotic map.
Key takeaways
- Reservoir computing trains only a readout layer; the reservoir (a recurrent network, a physical system, or living tissue) is left as it is.
- The classic forms are echo state networks (Jaeger) and liquid state machines (Maass, Natschläger and Markram, 2002).
- Biological neural tissue suits the role because it is nonlinear, recurrent and has fading memory of recent inputs.
- Brainoware tested speaker identification on 240 clips from eight speakers and Hénon map prediction; coverage reported about 78 percent speaker accuracy after training.
- It was less accurate than artificial neural networks on speech, and its authors called it a proof of concept, with challenges in organoid maintenance, interfaces and power.
What is reservoir computing?
Most neural networks learn by adjusting every connection weight through training. That works, but for recurrent networks (networks with loops, suited to time-varying data) it can be slow, unstable and expensive. Reservoir computing takes a shortcut.
A reservoir computer has three parts:
- An input layer that feeds a signal (audio, sensor readings, a time series) into the reservoir.
- A reservoir: a large, fixed, recurrent dynamical system. Inputs perturb it, and its internal state at any moment reflects both the current input and a fading trace of recent inputs.
- A readout: usually a linear model (such as linear or logistic regression) trained to map the reservoir’s state to the desired output.
Only the readout is trained. Tanaka and colleagues summarized the consequence in a 2019 Neural Networks review: because the reservoir is fixed and the readout uses simple methods such as linear regression, learning is fast and training costs are low. And because the reservoir needs no adaptive updating, it can be built from physical systems rather than software.
An intuition: dropping stones into a pond
A helpful picture is a pond. Drop in a stone and ripples spread, interfere and fade. Drop in a sequence of stones and the pattern on the surface at any instant encodes something about which stones fell, where and how recently. You do not need to control the water. You only need to learn how to read the ripples.
That picture is almost literal. Fernando and Sojakka (2003) built a reservoir from a container of water, in a paper titled “Pattern recognition in a bucket,” using the wave patterns as the reservoir’s state.
What makes a good reservoir?
Reservoir computing theory points to a few properties:
- Nonlinearity: the reservoir should transform inputs in nonlinear ways, so a linear readout can solve nonlinear problems.
- High dimensionality: many interacting elements, so different inputs map to distinguishable states (the “separation property”).
- Fading memory: the state should reflect recent inputs but gradually forget older ones (the “echo state” or fading-memory property).
- Consistency: the same input should produce similar responses, or the readout cannot learn a stable mapping.
How do echo state networks and liquid state machines work?
Reservoir computing emerged from two lines of work around the same time, which Lukoševičius and Jaeger later reviewed together in Computer Science Review (2009).
Echo state networks
Herbert Jaeger’s echo state network (ESN) uses a large, randomly connected recurrent network of simple artificial neurons as the reservoir. The connection weights are set randomly and scaled so that activity “echoes” inputs without exploding or dying out instantly. Only output weights are learned. In Science (2004), Jaeger and Haas reported that on a benchmark task of predicting a chaotic time series, ESNs improved accuracy by a factor of 2,400 over previous techniques. That paper brought wide attention to the approach.
Liquid state machines
At about the same time, Wolfgang Maass, Thomas Natschläger and Henry Markram proposed the liquid state machine (LSM) in Neural Computation (2002). Their reservoir, the “liquid,” is a recurrent circuit of spiking, integrate-and-fire neurons, closer to biology. They argued that the transient dynamics of a sufficiently large and heterogeneous neural circuit can serve as a universal analog fading memory, and that readout neurons can extract information about current and past inputs without the circuit needing stable internal states. The LSM was explicitly framed as a model of how cortical microcircuits might compute, which makes it a natural bridge to living tissue. It also connects to spiking neural networks and neuromorphic computing, which implement similar spiking dynamics in silicon.
Comparing reservoir types
| Reservoir type | Substrate | What is trained | Example |
|---|---|---|---|
| Echo state network | Software, random recurrent network of rate neurons | Linear readout only | Jaeger & Haas, 2004 |
| Liquid state machine | Software or hardware, spiking neuron circuit | Readout only | Maass, Natschläger & Markram, 2002 |
| Physical reservoir | Water, optics, electronics, spintronics, mechanics | Readout only | Fernando & Sojakka, 2003 (water) |
| Cultured neuron reservoir | Dissociated cortical neurons on an MEA | Readout only (tissue may also adapt) | Dockendorf et al., 2009 |
| Organoid reservoir | 3D brain organoid on a high-density MEA | Readout only (tissue may also adapt) | Cai et al., 2023 (Brainoware) |
Why does biological tissue fit reservoir computing?
If a reservoir needs to be nonlinear, high-dimensional, recurrent and to hold fading memory of recent inputs, a network of living neurons is an obvious candidate. A brain organoid contains many thousands of neurons of several types, connected in recurrent loops. Each neuron is strongly nonlinear: it either fires an action potential or it does not. Activity triggered by a stimulus reverberates for a while before settling. Those are exactly the properties the theory asks for.
There is also a practical attraction. In reservoir computing you never need to train the reservoir, which sidesteps the hardest problem in biocomputing: we do not know how to set the strength of every synapse in a living network. You only need to stimulate it consistently and read it well. The interface for that is a microelectrode array, which we explain in detail in microelectrode arrays explained.
Earlier work with cultured neurons
The idea predates organoids. Dockendorf and colleagues (2009, Biosystems) replaced the “liquid” of a liquid state machine with a living network of cortical neurons on an MEA. They applied high- and low-frequency multi-site stimulation patterns and used template-based classifiers to test whether the network’s responses kept inputs separable. They reported over 95 percent input reconstruction accuracy, rising to 99.8 percent when nonstationarity was compensated, confirming the separation property in these biological networks.
A twist: living reservoirs change
A digital reservoir stays fixed unless you change it. A living one does not. Neurons adapt, synapses strengthen and weaken through synaptic plasticity, and cultures develop over weeks. This cuts both ways:
- Potential benefit: stimulation might gradually shape the reservoir in useful directions, a kind of unsupervised learning that a fixed reservoir cannot do.
- Potential problem: if the reservoir drifts, a readout trained yesterday may perform worse today.
Brainoware’s authors reported evidence for the first effect, as described below. Managing the second is one of the field’s open engineering problems.
What is Brainoware and how was it set up?
Brainoware was described by Hongwei Cai and colleagues in “Brain organoid reservoir computing for artificial intelligence,” published in Nature Electronics in December 2023. The team, led by bioengineer Feng Guo at Indiana University Bloomington, combined a human brain organoid with a high-density MEA and conventional computing hardware. Nature Electronics published an accompanying commentary, “Reservoir computing with brain organoids,” by Lena Smirnova, Brian Caffo and Erik Johnson.
Based on coverage of the paper in MIT Technology Review, Tech Xplore and News-Medical, the setup had the classic three-part structure:
- Input layer. Time-dependent input data (for example, features from an audio clip) were converted into spatiotemporal sequences of electrical stimulation delivered through the MEA.
- Reservoir. A brain organoid grown from human pluripotent stem cells, containing multiple brain cell types and showing network electrical activity, sat on the high-density MEA. Stimulation perturbed its activity.
- Output layer. The organoid’s recorded responses were fed to a trained digital readout. For the speech task, coverage reported that this readout used logistic regression.
The organoid’s role is the one a random recurrent network plays in an echo state network: to turn inputs into rich, nonlinear, time-dependent patterns that a simple classifier can separate. How organoids are produced is covered in how brain organoids are grown, and HD-MEAs in our glossary entry on the high-density microelectrode array.
What tasks did Brainoware perform, and how well?
The study used two standard reservoir computing benchmarks.
Task 1: speaker identification from Japanese vowels
The first task was to identify who was speaking. Coverage described a dataset of 240 audio clips of eight male Japanese speakers pronouncing vowels. A Japanese vowels speaker-classification benchmark has long been used in this area; it traces back to a dataset introduced by Kudo, Toyama and Shimbo (1999, Pattern Recognition Letters).
The audio was converted into stimulation sequences, the organoid’s responses were recorded, and the readout learned to assign each clip to a speaker. Reported results:
- MIT Technology Review reported that after training, accuracy reached about 78 percent.
- News-Medical, summarizing the paper, reported that accuracy improved over training from 51 percent to 78 percent, which the authors attributed to unsupervised learning in the organoid.
- MIT Technology Review also reported that Brainoware was less accurate than artificial neural networks on this task.
For context, chance level with eight speakers is 12.5 percent, so 78 percent is well above chance but not state of the art for this kind of benchmark.
Task 2: predicting the Hénon map
The second task was prediction of a chaotic system. The Hénon map, introduced by Michel Hénon in Communications in Mathematical Physics (1976), is a simple two-dimensional rule that, with standard parameters, produces a chaotic “strange attractor.” Its next state depends nonlinearly on the current one:
- x(n+1) = 1 − a·x(n)² + y(n)
- y(n+1) = b·x(n)
With the classic values a = 1.4 and b = 0.3, the sequence never settles or repeats, yet it is fully deterministic, which makes it a standard test of whether a system can capture nonlinear temporal dynamics.
News-Medical reported that the two-dimensional map was reduced to a one-dimensional input for Brainoware, and that the system outperformed baselines including linear regression and an artificial neural network. Tech Xplore, reporting from Indiana University’s announcement, said Brainoware predicted the Hénon map with greater accuracy than an artificial network. Coverage did not report exact error figures, and summaries describe the neural network baselines differently, so the paper’s own figures are the place to check exactly which architectures it beat.
Controls and the evidence for learning
Two further findings from the coverage are worth noting:
- An organoid-free control, with no tissue on the array, produced a regression score of zero on the prediction task. In other words, the performance came from the tissue, not from the electronics or the readout alone.
- Training was reported to reshape the organoid’s functional connectivity, which the authors interpreted as unsupervised learning, and the effect was reported to depend on neural plasticity.
That last point is what distinguishes Brainoware from a pure reservoir. The tissue was not just a passive transformer of input. Repeated stimulation appeared to change it in a way that helped the readout. This echoes the closed-loop learning experiments described in how neurons learn in a dish, although Brainoware’s adaptation was driven by input exposure rather than by feedback about performance.
Brainoware at a glance
| Element | What was reported |
|---|---|
| Publication | Cai et al., Nature Electronics, December 2023 |
| Team | Indiana University Bloomington (lead researcher Feng Guo) |
| Reservoir | Human pluripotent stem cell-derived brain organoid |
| Interface | High-density multielectrode array |
| Input encoding | Time-dependent data converted to spatiotemporal stimulation sequences |
| Readout | Trained digital model (logistic regression for speech, per coverage) |
| Task 1 | Speaker identification, 240 clips, 8 male Japanese speakers |
| Task 1 result | About 78% after training, reported as up from 51%; below artificial neural networks |
| Task 2 | Hénon map prediction |
| Task 2 result | Reported to beat linear regression and an artificial neural network baseline |
| Stated status | Proof of concept |
How does Brainoware compare with digital reservoirs and ANNs?
It helps to be precise about what kind of comparison is meaningful.
Against digital neural networks
On accuracy, Brainoware is not competitive with modern machine learning. Digital models routinely do better on speaker identification, and MIT Technology Review reported Brainoware was less accurate than artificial neural networks on that task. Its reported advantage on the Hénon map applies to particular baselines, not to the best digital methods available.
Against digital reservoirs
The more natural comparison is with an echo state network or liquid state machine of similar size. Here the trade-offs are about properties, not just scores:
| Property | Digital reservoir (ESN or LSM) | Organoid reservoir (Brainoware) |
|---|---|---|
| Reservoir cost | Simulated on a CPU or GPU, essentially free to copy | Must be grown and kept alive continuously |
| Reproducibility | Identical copies with the same random seed | Each organoid differs, and it changes over time |
| Tuning | Spectral radius, size, sparsity adjustable | Limited control over structure |
| Adaptation | Fixed unless redesigned | May self-organize with stimulation (reported in Brainoware) |
| Interface | Direct memory access to every unit | Only what electrodes can stimulate and record |
| Energy | Simulation runs on conventional hardware | Tissue itself uses little power, but support hardware dominates today |
What about energy?
Energy efficiency is often cited as the long-term motivation for biocomputing, a case we examine in biocomputing vs silicon. But in Brainoware, as in every current system, the incubator, pumps, MEA electronics and computers use vastly more power than the organoid. Coverage noted that the authors themselves listed power use by peripheral equipment as a challenge. Any energy advantage is a future possibility, not a present result.
What are the limitations of Brainoware?
The authors and outside commentators were clear about the system’s limits. Drawing on the coverage, the main ones are:
- Organoid generation and maintenance. Organoids must be grown, fed and kept healthy around the clock. Tech Xplore’s report quoted Guo describing this as a 24/7 task.
- Variability. No two organoids are the same, which makes results harder to reproduce across samples.
- Flat, rigid electrodes. The MEAs used are planar and rigid, so they contact only part of a 3D organoid. Shell-shaped and mesh electrodes are being developed to address this.
- Peripheral power. Supporting hardware consumes far more energy than the tissue.
- Data tools. The authors pointed to a lack of efficient data management tools for this kind of experiment.
- Untested capabilities. MIT Technology Review noted the study did not show whether Brainoware can process and store information over the long term, or learn multiple tasks.
- No “hearing.” Lena Smirnova of Johns Hopkins University told MIT Technology Review that organoids cannot truly hear speech; they react to electrical pulses derived from the audio. All of the interpretation happens in the digital readout.
Guo described the study as “just proof-of-concept” and said it may be decades before general biocomputing systems can be created.
What open questions remain for organoid reservoir computing?
Brainoware showed that an organoid can serve as a working reservoir. Turning that into a useful computing approach raises questions that remain open:
- Stability vs plasticity. How do you get the benefits of self-organization without the readout drifting out of date? Periodic readout recalibration is one option; controlling plasticity is another.
- Encoding. What is the best way to translate data into stimulation? Choices about which electrodes, what pulse rates and what timing likely matter as much as the tissue. This is a core problem in neural encoding and decoding.
- Memory capacity. How long does an organoid “remember” an input, and how does that depend on age, size and cell composition?
- Scaling and access. Larger, more mature organoids, supported by microfluidics and organoid vascularization, might offer richer dynamics, but only if interfaces can reach them.
- Multiple organoids. Could several organoids, or connected assembloids, work as larger or modular reservoirs?
- Benchmarks and baselines. Fair comparisons need size-matched digital reservoirs, standard tasks and shared data, ideally in open formats such as Neurodata Without Borders.
- Ethics. As organoids become more complex, questions about consent for donated cells and the moral status of tissue become more pressing. We discuss these in the ethics of organoid intelligence.
The organoid intelligence roadmap by Smirnova and colleagues (2023, Frontiers in Science) frames many of these as priorities for the field as a whole, from 3D interfaces to data infrastructure. Reproducibility depends on recording exactly how each organoid was stimulated and read out, which is the problem we’re working on at Organoid Programming (our platform).
Conclusion
Reservoir computing with brain organoids uses living tissue as a fixed, nonlinear, fading-memory system and trains only a simple readout, the same idea behind echo state networks and liquid state machines. Brainoware put that idea to work: an organoid on a high-density MEA helped identify eight speakers from Japanese vowel clips with about 78 percent accuracy after training and was reported to beat simple baselines at predicting the chaotic Hénon map. It did not beat modern neural networks, and its authors called it a proof of concept. Its significance lies in showing that organoid intelligence research can use standard machine learning benchmarks, and in exposing the engineering problems that come next. For the broader picture, see what organoid intelligence is and the future of biocomputing.
Frequently asked questions
What is reservoir computing?
Reservoir computing is a machine learning approach in which input signals drive a fixed, complex dynamical system called the reservoir, and only a simple readout layer is trained to map the reservoir’s states to outputs. Because the reservoir itself is never trained, learning is fast and cheap. Echo state networks and liquid state machines are the classic examples.
What is Brainoware?
Brainoware is a system described by Cai et al. in Nature Electronics in 2023, from a team at Indiana University Bloomington. A human brain organoid on a high-density microelectrode array served as the reservoir. Inputs were converted into patterns of electrical stimulation, the organoid’s responses were recorded, and a trained digital readout produced the final output.
What tasks did Brainoware perform?
The study tested two benchmark tasks: identifying which of eight speakers produced a recording, using 240 audio clips of Japanese vowel sounds, and predicting the Hénon map, a classic chaotic dynamical system. Coverage of the paper reported speaker identification accuracy of about 78 percent after training. The authors described the work as a proof of concept.
Why use a brain organoid as a reservoir?
A good reservoir needs many interacting nonlinear elements, rich dynamics and fading memory of recent inputs. Living neural tissue has all three: thousands of neurons, recurrent connections, nonlinear firing and activity that echoes for a short time after stimulation. It may also adapt through plasticity, which is unusual for a reservoir and could be either a benefit or a complication.
Is Brainoware better than artificial neural networks?
Not in general. MIT Technology Review reported that Brainoware was less accurate than artificial neural networks on the speech task. Coverage reported it did better than some simple baselines on the Hénon map task. The study’s value is as a demonstration that organoid tissue can act as a working reservoir, not as a competitive replacement for digital machine learning.
Did the brain organoid hear or understand speech?
No. The organoid received electrical stimulation patterns derived from audio features, not sound. Lena Smirnova of Johns Hopkins University told MIT Technology Review that organoids cannot truly hear speech and only react to the stimulation pulses. The classification was done by a trained digital readout using the organoid’s responses.
What are the main limitations of organoid reservoir computing?
Organoids require continuous care, vary from one sample to another, and change over time. Flat electrode arrays sample only part of the tissue, the supporting hardware uses far more power than the tissue itself, and data tools are immature. Brainoware also did not test long-term information storage or learning multiple tasks.
Sources
- Cai et al. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics.
- Smirnova, Caffo & Johnson (2023). Reservoir computing with brain organoids. Nature Electronics.
- MIT Technology Review (2023). Human brain cells hooked up to a chip can do speech recognition.
- News-Medical (2023). Brainoware: a breakthrough AI approach using brain organoids for advanced computation.
- Tech Xplore (2023). Brain tissue on a chip achieves voice recognition (Indiana University).
- Jaeger & Haas (2004). Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science.
- Maass, Natschläger & Markram (2002). Real-time computing without stable states: a new framework for neural computation based on perturbations. Neural Computation.
- Lukoševičius & Jaeger (2009). Reservoir computing approaches to recurrent neural network training. Computer Science Review.
- Tanaka et al. (2019). Recent advances in physical reservoir computing: a review. Neural Networks.
- Fernando & Sojakka (2003). Pattern recognition in a bucket. In Advances in Artificial Life (Springer).
- Dockendorf et al. (2009). Liquid state machines and cultured cortical networks: the separation property. Biosystems.
- Hénon (1976). A two-dimensional mapping with a strange attractor. Communications in Mathematical Physics.
- Kudo, Toyama & Shimbo (1999). Multidimensional curve classification using passing-through regions. Pattern Recognition Letters.
- Smirnova et al. (2023). Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science.
